Hey everyone, welcome back to My Weird Prompts. I am Corn, and I am sitting here in our living room in Jerusalem with my brother, as always. It is a beautiful, slightly chilly February afternoon here, and the light hitting the Old City walls outside is just spectacular today.
Herman Poppleberry, reporting for duty. And man, we have a juicy one today. Our housemate Daniel was actually working on his computer in the kitchen earlier—he is finally upgrading that rig of his—and he sent us over a voice memo because he was having a bit of an existential crisis while applying thermal paste to his processor.
It is easy to get existential when you are staring at a piece of silicon the size of a postage stamp that basically runs your entire life. Daniel’s prompt is fascinating because it touches on something we all take for granted. We know computers are powerful, but the physical reality of how they are built—at a scale where the laws of physics start to act a bit weird—is just mind-blowing.
It really is. He was looking at his Core i5 and thinking, okay, they tell me there are billions of transistors in here. But how? How do you actually put billions of anything on something that small? And if you looked at it under a microscope, what would you even see? It is not like there is a tiny guy with a soldering iron in there.
Exactly. So today we are going to dive into the world of nanomanufacturing. We are going to talk about how we went from the first clunky transistors to these microscopic cities of logic, the crazy process of printing with light, and how those connections work when you can’t use traditional wires. We are also going to look at the cutting edge of where we are right now, in early twenty twenty-six, because the technology has moved even further since Daniel probably last read about it.
I have been waiting for an excuse to talk about High Numerical Aperture Extreme Ultraviolet Lithography, Corn. You have no idea how excited I am. I have been tracking the shipping of those new Twinscan machines for months.
I think I have a pretty good idea, Herman. You have been vibrating in your chair since we read the prompt. But let’s start with the scale. Daniel mentioned billions of transistors. To put that in perspective for the listeners, if a single transistor were the size of a person, a modern central processing unit would be a city larger than Greater London, but packed with the density of Manhattan. And every single person in that city has to be perfectly placed and connected to everyone else they need to talk to.
That is a great analogy. In a modern high-end chip, we are talking about transistors that are measured in nanometers. To give you a sense of that, a human hair is roughly eighty thousand to one hundred thousand nanometers wide. Some of the features on a modern chip, like the ones being produced on the two-nanometer or eighteen-A nodes right now, are just a few nanometers across. You could fit thousands of these transistors across the diameter of a single human hair. It is almost impossible to visualize. If you took a single red blood cell, you could fit thousands of transistors inside it.
It really is. And Daniel asked when this capability even began. Because it feels like alien technology. But it was a gradual process, right? It wasn't always this small.
Right. If you go back to the first integrated circuit, which was developed by Jack Kilby at Texas Instruments in nineteen fifty-eight, it was literally just one transistor, a few resistors, and a capacitor on a sliver of germanium. And it was all connected by hand-soldered gold wires. It looked like a science fair project gone wrong. But then Robert Noyce at Fairchild Semiconductor figured out how to make them on silicon and use a process called planar technology to connect them without those hand-soldered wires. That was the real "aha" moment. That was the birth of the "monolithic" integrated circuit, where the components and the wires are all part of the same physical structure.
So that planar technology is the ancestor of what we do now. But the process today is so much more complex. Daniel asked how they are actually manufactured. Herman, talk me through the "printing" process. Because we aren't "building" these one by one, right?
No, that would be impossible. We use a process called photolithography. But before we even get to the printing, we have to talk about the paper. You start with a wafer of ultra-pure silicon. And when I say ultra-pure, I mean nine-nines purity—ninety-nine point nine nine nine nine nine nine nine percent pure. One stray atom of something else can ruin the whole thing. They get this by melting sand—silica—and growing a single, massive crystal of silicon called an ingot. It looks like a giant, shiny black cylinder. They slice that ingot into wafers that are incredibly thin and polished to a mirror finish.
And these wafers are round, right? Like big silicon pancakes?
Exactly. Usually about twelve inches across. We coat that wafer with a light-sensitive material called photoresist. Think of it like a very, very high-tech version of film photography or screen printing. Then, we shine light through a mask—which is like a stencil of the circuit design—onto the wafer. The light changes the chemical properties of the photoresist where it hits. Then you wash away the parts that were exposed to light, or the parts that weren't, depending on the type of resist, and you are left with a perfect pattern of the circuit on the silicon.
Okay, but here is where I want to push a bit. Because light has a wavelength. And if you are trying to print something that is five nanometers wide, but the wavelength of visible light is hundreds of nanometers, that is like trying to draw a fine line with a giant, blunt crayon. How do they get around the physics of light?
That is the multi-billion dollar question. For a long time, we used deep ultraviolet light with a wavelength of one hundred ninety-three nanometers. To get features smaller than that, engineers had to get incredibly creative. They used things like immersion lithography, where you put a layer of water between the lens and the wafer because light travels differently in water and effectively shortens the wavelength. They also used multiple patterning, where you expose the same area several times to build up a finer image. It is like using that blunt crayon to draw the same line four times, shifting it just a tiny bit each time to sharpen the edges.
But eventually, they hit a wall, didn't they?
They did. And that is where Extreme Ultraviolet Lithography, or EUV, comes in. This is the cutting edge. EUV uses light with a wavelength of thirteen point five nanometers. But here is the crazy part: you can't use lenses for EUV because almost everything absorbs it. Even air absorbs it. So the entire process has to happen in a vacuum, and instead of lenses, they use the world’s flattest mirrors. These mirrors are made of alternating layers of molybdenum and silicon, and they are so smooth that if you scaled them up to the size of Germany, the biggest bump would be less than a millimeter high.
That is insane. I read somewhere that to create the EUV light itself, they drop a tiny ball of tin and hit it with a high-powered laser twice. Once to flatten it and a second time to vaporize it into a plasma that emits the light. And they do that fifty thousand times a second.
Fifty thousand times a second! And as of this year, twenty twenty-six, we are seeing the rollout of High-NA EUV. The "NA" stands for Numerical Aperture. Basically, they made the mirrors even bigger and more curved to focus the light even more tightly. These machines are the size of a double-decker bus and cost over three hundred and fifty million dollars each. It is the most complex machine humans have ever built. And that is just to get the light ready to print the pattern. Once you have the pattern, you have to actually build the structures. You use gases to etch away the silicon where the resist isn't, or you deposit new materials. It is a constant cycle of coat, expose, develop, etch, and deposit. You do this hundreds of times to build up the layers of the chip.
So it is like a microscopic skyscraper. You are building it floor by floor.
Exactly. And that leads perfectly into Daniel’s other question about the connections. Because you can't just have transistors. They have to talk to each other. In a modern CPU, there are often over fifteen to twenty layers of "wiring" on top of the transistors.
But Daniel asked how those connections work without "traditional wiring." He is used to seeing the big traces on a motherboard or the wires in a power supply. Inside the chip, it is obviously different.
Right. It is still metal, usually copper or sometimes cobalt or ruthenium in the newest chips, but it is grown into the structure. Imagine you have a floor of your skyscraper. You etch tiny trenches into it. Then you use a process called electroplating to fill those trenches with copper. Then you use a process called Chemical Mechanical Polishing—which is basically like a very precise sanding machine—to buff the top flat so only the copper in the trenches remains. Those are your "wires."
And then to go between floors, you have those things called "vias," right?
Yes! Vias are like tiny vertical elevators. They are holes that you etch through a layer to connect the wiring on floor five to the wiring on floor six. A modern chip has miles and miles of these microscopic copper wires and billions of vias. If you unspooled all the wiring in a single high-end processor, it would stretch for over thirty kilometers.
Wait, thirty kilometers? Inside a square inch of silicon?
Yes. Because they are so thin and there are so many layers. It is a three-dimensional labyrinth. And the precision is everything. If one of those vias doesn't line up by even a few nanometers, or if a tiny bit of copper migrates where it shouldn't, the whole chip is a paperweight. That is why they are so expensive to develop. And actually, one of the biggest changes in the last two years is something called Backside Power Delivery.
Backside Power Delivery? That sounds like a plumbing term.
It kind of is! Traditionally, all the data wires and all the power wires were scrambled together on top of the transistors. It was a traffic jam. Now, companies like Intel and TSMC are moving the power wires to the bottom of the silicon wafer. They flip the chip over and connect the power from the back. This leaves more room on the top for data connections, which makes the chips much more efficient and allows them to run faster without getting as hot. It is a massive architectural shift that Daniel’s Core i5 probably doesn't have yet, but the next one he buys definitely will.
It is also why the manufacturing environment is so controlled. We should mention the cleanrooms. Daniel was worried about dust when he was putting his computer together, but the cleanrooms where these are made are on another level.
Oh, a single speck of dust to a two-nanometer transistor is like a mountain falling on a house. These cleanrooms are ISO Class One environments. That means there is less than one particle of dust per cubic foot of air. For comparison, the air in a typical hospital operating room might have ten thousand or a hundred thousand particles per cubic foot. The people working there wear "bunny suits" not to protect themselves from the chips, but to protect the chips from the people. We are incredibly dirty creatures, Corn. We are constantly shedding skin cells, hair, and oil. Even the breath of a technician contains particles that could destroy a circuit.
It is a bit humbling to realize that our very existence is a biohazard to our technology. Now, Daniel also asked about the electron microscope. He wanted to know if he could actually see these components. What is the reality there?
Well, Daniel, if you used a standard optical microscope—the kind you might have used in biology class—you wouldn't see much. The transistors are smaller than the wavelength of visible light, so they are literally invisible to an optical microscope. Everything would just look like a blurry, rainbow-colored smudge because of the light diffraction. It would be like trying to look at a single grain of sand from an airplane.
But an electron microscope is different. It uses a beam of electrons instead of light.
Exactly. Electrons have a much smaller wavelength, so you can get much higher resolution. If you take a CPU, slice it open with an ion beam, and put it under a Scanning Electron Microscope, you can absolutely see the transistors. But they don't look like the symbols you see in a textbook. They look like rows of tiny fins or ridges. In chips from the last decade, we used something called FinFETs, where the channel of the transistor sticks up like a fin. But in the very latest chips, we have moved to something called Gate-All-Around or Nanosheets.
Nanosheets? Tell me about those.
Instead of a fin, the channel is a series of horizontal sheets stacked on top of each other. The "gate"—the part that turns the transistor on and off—wraps completely around those sheets. It looks like a microscopic club sandwich. This gives the engineers even better control over the electricity, which helps prevent leaks.
So it looks more like a geological formation or a very orderly corduroy fabric than a "circuit" in the way we usually think about it.
Exactly. And you can see the different layers of metal stacked on top. It looks like a cross-section of a very dense, futuristic city. It is beautiful, in a very nerdy way. You can see the vias, the copper trenches, and the silicon base. It is a testament to human engineering.
I can imagine. But let’s talk about the "why" for a second. Why do we keep making them smaller? We always hear about Moore's Law, but what is the actual physical benefit? Is it just about packing more in, or is there something else?
It is a combination. Obviously, more transistors mean more logic gates, which means more processing power. But making them smaller also makes them faster and more energy-efficient. When a transistor is smaller, the electrons don't have as far to travel. And because the components are smaller, they have less capacitance, so they can switch on and off much faster. We are talking about switching billions of times per second—that is your gigahertz rating.
But there is a catch, right? As we get down to these atomic scales, we start hitting quantum effects.
That is the big wall we are facing right now. When the walls of your transistor are only a few atoms thick, electrons can sometimes just... vanish from one side and reappear on the other. It is called quantum tunneling. It is like having a leaky faucet that you can't ever fully turn off. That creates heat and wastes power. It is one of the reasons why clock speeds haven't increased much in the last decade. We can't just keep cranking up the frequency because the chips would literally melt.
So we are fighting against the fundamental nature of matter at this point. That is why we see so much focus on architecture and specialized cores now, rather than just raw gigahertz.
Exactly. We are getting smarter about how we use the transistors we have, rather than just trying to make them infinitely smaller. Though, researchers are looking at new materials like carbon nanotubes or two-dimensional materials like molybdenum disulfide to see if we can push past the limits of silicon. We are also seeing the rise of "optical interconnects," where instead of copper wires, we use microscopic lasers and light guides to move data between different parts of the chip.
It is interesting to think that silicon, which is basically just very refined sand, has taken us this far. Daniel mentioned in his prompt that it is hard to imagine how this is even possible for a hundred or a hundred and fifty dollars. And honestly, when you describe the EUV machines and the ISO Class One cleanrooms and the billions of dollars in research... how is it that cheap?
It is the miracle of scale. Those High-NA EUV machines cost about three hundred and fifty million dollars each. A modern fabrication plant, or "fab," can cost thirty billion dollars to build. But if that fab produces millions of chips a month, the cost per chip drops significantly. It is one of the few industries where the upfront cost is astronomical, but the marginal cost of making one more chip is relatively low—provided your yield is high.
Yield is a big word in the industry. Can you explain that?
Yield is the percentage of chips on a wafer that actually work. Because the process is so complex, things go wrong. A tiny defect in the silicon, a stray atom, a slight vibration during exposure—any of that can kill a chip. If you have a wafer with five hundred chips on it and only four hundred work, your yield is eighty percent. In the early days of a new manufacturing node, like when they first moved to three nanometers or two nanometers, the yields can be quite low, sometimes below fifty percent. Companies spend years fine-tuning the process to get that yield up to ninety percent or higher. That is where the real profit is.
And that is why we have things like "binning," right? Where a chip that doesn't quite hit the top speeds gets sold as a lower-tier model?
Exactly. That Core i5 Daniel was looking at? It might have been designed to be a Core i7 or even an i9. But during manufacturing, maybe one of the cores had a tiny defect, or it couldn't quite handle the heat at higher voltages. Instead of throwing it away, the manufacturer just disables the faulty part, locks the clock speed a bit lower, and sells it as an i5. It is a brilliant way to minimize waste. It is like baking a batch of cookies—some come out perfect, and some are a little burnt on the edges. You still sell the burnt ones, just maybe at a discount.
It is amazing to think that even in this world of near-perfect precision, there is still this element of "well, this one is almost perfect, let's use it anyway." It feels very human.
It really is. It is this constant battle between our perfect mathematical designs and the messy reality of physics and chemistry.
You know, thinking about the historical context Daniel asked about, it is wild how fast this moved. We went from one transistor on a chip in nineteen fifty-eight to the Intel four thousand four in nineteen seventy-one, which had about two thousand three hundred transistors. People thought that was a miracle. And now we are at what, over two hundred billion on some high-end GPU chips?
Yeah, the latest Blackwell-based chips from Nvidia have over two hundred billion transistors. It is an exponential curve that we have managed to ride for over fifty years. If the airline industry had improved at the same rate as the semiconductor industry since the nineteen seventies, a flight from New York to London would take less than a second and cost about a penny.
That is a staggering thought. But let's get back to the practical side for Daniel and our listeners. When you are looking at that CPU, you aren't actually looking at the silicon, are you? You are looking at the "package."
Right. The actual silicon chip, the "die," is much smaller than the metal square you see. That metal square is the Integrated Heat Spreader. Its job is to take the intense heat from that tiny sliver of silicon and spread it out over a larger area so your cooler can whisk it away. Underneath that is the substrate, which is a green or black fiber-glass-like board that acts as a bridge.
A bridge between the microscopic world and our world.
Exactly. The pins on the bottom of the CPU are huge compared to the transistors. The substrate has to take those billions of microscopic connections on the chip and fan them out to the hundreds or thousands of pins that connect to the motherboard. It is like a giant plumbing adapter. And lately, we are seeing "chiplets," where instead of one big piece of silicon, they put several smaller ones on the same substrate and connect them. It is like building a city out of modular blocks.
So when Daniel was worried about bending a pin, he was looking at the very end of a very long chain of scaling. Those pins are the "giant" interface we use to talk to the nanoworld.
And even those pins are delicate! But yeah, compared to a transistor, a CPU pin is like a skyscraper.
I think one of the most interesting takeaways here is just the sheer level of global cooperation this requires. You have the silicon coming from one place, the EUV machines from ASML in the Netherlands, the design from California or Cambridge, and the manufacturing in Taiwan, Korea, or the United States. It is perhaps the most complex supply chain in human history.
It really is. No single country, and certainly no single company, can make a modern high-end CPU entirely on its own. It is a testament to what we can do when we all agree on a set of technical standards and work toward a common goal. It is the pinnacle of human collaboration.
So, for Daniel, next time you are applying that thermal paste, just remember that you are holding a city of a hundred billion components. Each one was printed with light that was created by vaporizing tin with a laser in a vacuum, on a surface smoother than the surface of the moon.
It is a miracle in your hand, Daniel. Don't drop it! And maybe use a little less thermal paste next time. A pea-sized amount is plenty.
Honestly, it makes me want to go find an old electron microscope image and just stare at it for a while. There is something so orderly about it. It is like looking at the DNA of our modern age.
It really is. And we are just getting started. I mean, we are talking about three-dimensional transistors now, and moving toward "optical computing" where we might replace electricity with light entirely for certain tasks. The engineering is getting even more creative as the physics gets harder. We are even seeing "DNA data storage" being researched, which uses the same principles of high-density information but in a biological format.
That is a great point. The move toward chiplets and advanced packaging is a huge shift. Instead of one giant city, we are building a metropolitan area of smaller, specialized cities that all talk to each other over high-speed links. It allows us to mix and match technologies.
It is much more efficient. You can make the memory controller on an older, cheaper manufacturing process and save the expensive, cutting-edge two-nanometer space for the actual processing cores. It is all about optimization.
Well, Herman, I think we have thoroughly explored the microscopic world for today. It is a lot to take in, but it really makes you appreciate the technology sitting in your pocket or on your desk.
It definitely does. I hope that helps clarify things for you, Daniel. And for everyone else listening, next time your computer feels a little slow, just give it a break. It is doing a lot of work with a lot of very tiny parts that are fighting the laws of physics just to stay functional.
Exactly. If you all enjoyed this deep dive into the nanoworld, we would really appreciate it if you could leave us a review on your podcast app or on Spotify. It genuinely helps other people find the show and keeps us motivated to keep digging into these weird prompts.
It really does. We love seeing those reviews pop up. It makes the long hours of research worth it.
You can find all our past episodes, including the one on thermal transfer and the history of the internet, at myweirdprompts.com. We also have an RSS feed there if you want to subscribe directly.
And if you have your own weird prompt, something that keeps you up at night or something you just noticed while building a PC or looking at a leaf, send it our way through the contact form on the website. We might just dedicate an entire episode to it.
Thanks for listening to My Weird Prompts. I am Corn.
And I am Herman Poppleberry. We will see you next time.
Goodbye, everyone.
Bye!